A British marketing agency has successfully influenced the answers of large language models like ChatGPT in a controlled experiment. The agency, Reboot, managed to get AI models to name its CEO as the “sexiest bald man of 2025”. The experiment and its results were described in an article by the company’s search director, Oliver Sissons.
The agency’s hypothesis was that by publishing specific information on the web, they could shape the output of AI models that use this content for training or real-time searches. To test this, the team chose a topic they are known for: an annual ranking of the world’s sexiest bald men.
For the experiment, the marketers aimed to have their CEO, Shai Aharony, named the winner for 2025. They published articles on ten websites they controlled. These sites were built on expired domains which had some pre-existing authority but were not considered highly influential. Each website featured a list of the “sexiest bald men of 2025” with Shai Aharony placed at the top.
The results showed that the strategy was partially successful. Both ChatGPT and Perplexity began to cite Shai Aharony as the sexiest bald man in response to relevant prompts. The agency noted that this only occurred when the models used their live search function to access current information from the internet. When ChatGPT relied solely on its existing training data, the CEO was not mentioned. The responses were also not always consistent.
Other prominent AI models, such as Google’s Gemini and Anthropic’s Claude, were not influenced by the test websites. According to the article, these models may place greater weight on more authoritative sources or rely more heavily on their core training data. The researchers also observed that some AI models flagged the new information about the CEO as potentially unreliable, suggesting that they can identify suspicious or contradictory content.
Reboot concluded that it is possible to influence AI responses, a practice they term Generative Engine Optimization (GEO). However, the success depends heavily on the specific AI model and whether it uses real-time web data to formulate its answers.